Prune uneven 4A/4B partitions with neighbor block boundaries

[urvangjoshi]

Use the best partitioning boundaries of left / above coding blocks to prune uneven 4way partitions -- by checking boundary alignment.

Only prune for inter frames.

• Enabled for speed >= 1.
• But disabled for resolution <= 270p.

Results for RA CTC 33 frames, speed 1:
(Anchor: 8d95884)

+---------+--------+--------+--------+--------+----------+
| Summary |   Y    |   U    |   V    |  YUV   | Enc-time |
+---------+--------+--------+--------+--------+----------+
| A1      |  0.08% |  0.04% |  0.11% |  0.08% | 97.10%   |
| A2      |  0.07% | -0.07% |  0.16% |  0.07% | 96.50%   |
|avg wo b2|  0.07% |  0.09% |  0.06% |  0.07% | 96.75%   |
+---------+--------+--------+--------+--------+----------+

Note: no changes at speed 4, because uneven 4way partitions are disabled at this speed.

Related: refactor a common av2_get_chroma_start_location() function.

STATS_CHANGED for speed >= 1.

[leolzhao]

The trade off for class A2 is pretty good, however the trade off for class A1 is worse than the threshold. Currently, our rule is to make sure the trade off for both class A1 and A2 should satisfy the threshold. So, is it possible to add the qp dependent condition for this speed feature? For example, only turn on this speed feature for 4K sequence when qp is greater than one threshold? @urvangjoshi

[urvangjoshi]

The trade off for class A2 is pretty good, however the trade off for class A1 is worse than the threshold. Currently, our rule is to make sure the trade off for both class A1 and A2 should satisfy the threshold. So, is it possible to add the qp dependent condition for this speed feature? For example, only turn on this speed feature for 4K sequence when qp is greater than one threshold? @urvangjoshi

Thanks @leolzhao for the suggestion! Let me try some variations in this direction.

[yeqing-wu]

Github uses markdown style in its conversation messages. To make sure that the results are not messed up by Github, please add ``` before and after the results.

[urvangjoshi]

The trade off for class A2 is pretty good, however the trade off for class A1 is worse than the threshold. Currently, our rule is to make sure the trade off for both class A1 and A2 should satisfy the threshold. So, is it possible to add the qp dependent condition for this speed feature? For example, only turn on this speed feature for 4K sequence when qp is greater than one threshold? @urvangjoshi

@leolzhao : I made some tweaks to the code as follows:

• Rewrote the code to remove some checks and add some higher level checks to avoid some computations.
• Disabled this speed feature for <= 270p resolution, as I was seeing a large loss on a5 set only (~0.3%).
• Reran on a cluster with more reliable timing.

With that, all sets seem to have a good trade-off now. See the updated commit message for latest results.

[urvangjoshi]

@leolzhao : I'm not sure why the CI run is failing in "common variable setup" phase. Seems to be a transient issue perhaps.
You may need to force-merge the PR.